Menu
Artificial Intelligence has been applied in many aspects of life, however, AI algorithms have been less used in the field of music. In this paper, a multi-track based pop music generation model MuseGAN is proposed, due to its poor contextualization and excessive tempo jumps in generating pop music samples. In this paper, a new multi-track pop music generation model-Recurrent Feature Generation Adversarial Network RFGAN is proposed. the model addresses the temporal relevance of the music structure and the repetitive nature of the musical section, and proposes a temporal model that enhances the contextual relevance of the music samples in terms of the time series, and improves the generative model according to this temporal model by converting the unidirectional structure in the original model to a recurrent structure, adding the feature extractor to the previous level of training information, which is combined with arbitrary noise and passed to the next training. An average pooling layer is added at the end of the generative model as a solution to the situation where the model generates too much noise for pop music samples. The improved model is superior to the pre-improvement model in terms of stability, convergence speed, and overfitting in pop music generation. In the audience scoring experiment, 60% of the top 5 pop music scores were generated using the RFGAN model proposed in this paper, indicating that the pop music generated using the RFGAN model has reached a high level comparable to the level of artificial pop music composition.